The currently described disclosure relates to managing a prefetch buffer with probabilistic access predictions. Storage systems usually “prefetch” data from one or more disks for temporary storage in memory, e.g., Dynamic Random Access Memory (DRAM), when the system perceives that the pre-fetched data might be requested in the near future. The method used to predict which data will be needed is imperfect and usually relies on an extrapolation from observed data accesses by a user. In addition, the total amount of memory available to store such data is limited. Hence, when a new access is identified or predicted, the system necessarily must determine whether to prefetch additional data, and if so, which data currently in memory to evict to make space for storing the newly prefetched data.